When you connect a locally-served Hugging Face model to an AI agent app on your Mac, your custom endpoint may receive zero traffic, even though the chat appears to work normally. This silent failure stems from a model-selection routing bug that only affects Claude-compatible endpoints, not Gemini or Codex models. The issue has a clear cause and a straightforward fix, both documented in June 2026 releases.

What Happened in Hugging Face's June 2026 Model Releases?

Hugging Face added five new model classes to its Transformers library in June 2026, according to the official release notes. These additions span multiple AI domains, from text generation to speech recognition and document analysis.

• DiffusionGemma: An encoder-decoder text model optimized for fast inference using multi-canvas sampling, which denoises entire token blocks simultaneously rather than generating left to right.
• DeepSeek-V3.2-Exp: An experimental release introducing DeepSeek Sparse Attention (DSA), a trainable sparse attention mechanism designed for long-context training and inference efficiency.
• MiniMax-M3-VL: A vision-language model combining a CLIP-style vision tower with the MiniMax-M3 text backbone, using mixed dense and sparse Mixture-of-Experts architecture.
• PP-OCRv6: A lightweight optical character recognition system built from MetaFormer-style blocks, available in three tiers (medium, small, tiny) for document text recognition.
• Parakeet-RNNT: A speech-to-text model pairing a Fast Conformer encoder with an RNN-T decoder using LSTM prediction networks for language context.

The June 12 release of MiniMax-M3-VL, PP-OCRv6, and Parakeet-RNNT required immediate stabilization. Two patch releases landed on June 15, 2026, addressing vLLM compatibility issues, processor token ID problems, InternVL handling, PEFT lower bounds, and Mistral tokenizer backend issues.

Why Does Your Custom Endpoint Receive No Requests?

The routing failure occurs because of how macOS AI agent applications decide which backend to use. When you paste a custom endpoint URL into an app's settings, that URL is stored as the ANTHROPIC_BASE_URL environment variable. However, this variable only controls traffic on the Claude API path. The app determines which path to use based on your selected model: anything that is not a Gemini model and not a Codex model gets routed to Claude. But here is the catch: if your app defaults to Gemini Flash (which many do for speed and cost), your custom endpoint is completely bypassed, and your bridge logs remain empty.

This is not a misconfiguration on your end. It is a design flaw in how the model selector and endpoint interaction work. The app silently routes your request to Gemini instead of your local bridge, with no error message or warning to alert you that your custom endpoint was never used.

How to Fix the Silent Routing Bypass?

The fix requires two steps, both simple but easy to miss.

• Update Your App: Version 2.9.68, released on June 16, 2026, added a warning when your selected model is not a Claude model. The app now alerts you that your custom endpoint only routes Claude traffic and offers a one-tap switch to a Claude model. If you are running an older version, update immediately.
• Format Your Endpoint URL Correctly: Your endpoint must be an absolute HTTP or HTTPS URL with a host. A bare localhost:8766 with no scheme will be silently rejected and cause every query to fail with an invalid-URL error that the retry-with-resume path can swallow, resulting in an empty turn with no error message. Always write the full http://localhost:8766 format.
• Switch to a Claude Model: After updating, explicitly select a Claude model from the dropdown instead of relying on the default Gemini Flash. This ensures your requests actually reach your custom endpoint and your bridge logs show incoming traffic.

The root cause of this footgun is clear. Version 2.9.60, shipped on June 3, 2026, changed new chats to start on Gemini Flash by default because it is fast and free. That is a sensible default for most users, but it creates a hidden trap for anyone trying to use a custom endpoint. Two weeks later, version 2.9.68 closed the gap with a warning and a one-tap fix.

What Does This Mean for Running Local Models?

Connecting a local model to an agent app is not the same as making that model work well in agent workflows. Agent loops are demanding: they require long chains of tool calls, exact argument formatting, and complex context management. A model that works fine for simple chat may fail silently or produce malformed tool calls when used as an agent backend. The routing bug compounds this challenge by hiding whether your requests are even reaching your local model or being silently rerouted to a cloud API.

For developers and power users experimenting with local Hugging Face models in June 2026, the lesson is clear: verify your routing explicitly. Check your bridge logs to confirm traffic is arriving. Update to the latest app version. Format your endpoint URL correctly. And switch to a Claude model to ensure your custom endpoint is actually being used. Without these steps, you may spend hours debugging a model that is working fine, when the real problem is that your requests never reached it at all.